[petsc-users] Preconditioner question ASM vs SOR

[Matthew Knepley]

On Thu, Sep 17, 2015 at 7:39 AM, Erik Andreassen <erand at mek.dtu.dk> wrote:

> I think I understand; does this mean that it is possible to change the sub
> PC type to SOR? Or is there a specific reason why the sub type is ILU?
>

Yes, this is exactly it. We have a default of ILU because it works better
in a general purpose problem, but it usually not the solver you want to use
for any given problem.

  Thanks,

     Matt


> The reason why I am asking, is that when I look at the residuals coming
> out of PCMG, they seem to be largest on local domain boundaries
> (corresponding to the partitions). Therefore, I’d like to use ASM with
> overlap instead of SOR as the preconditioner on the coarser levels (to
> smooth these errors out), but with the ASM/ILU combination this does not
> work at all (in accordance with your comment on high frequencies).
>
>
>
> Thanks again,
>
> Erik
>
>
>
> *From:* Matthew Knepley [mailto:knepley at gmail.com]
> *Sent:* 17. september 2015 14:12
> *To:* Erik Andreassen
> *Cc:* petsc-users at mcs.anl.gov
> *Subject:* Re: [petsc-users] Preconditioner question ASM vs SOR
>
>
>
> On Thu, Sep 17, 2015 at 7:07 AM, Erik Andreassen <erand at mek.dtu.dk> wrote:
>
> I have played a bit with preconditioners for the different levels in PCMG,
> and among others I have tried additive Schwartz (ASM), which does not seem
> to work very well – especially with no overlap. I’m curious to hear if
> anyone can explain why ASM with no overlap performs so much worse than SOR?
>
>
>
> The point of a smoother is to wipe out the high frequencies in the error.
> SOR does this (provably for the Laplacian), but there is no
>
> reason ASM/ILU should do this. If this does not happen, then these parts
> of the solution remain uncorrected when you go to the
>
> coarse grid, and the convergence breaks down.
>
>
>
>    Matt
>
>
>
> Thanks,
>
> Erik
>
>
>
> PS: I had a previous question about setting levels in GAMG, and my first
> impression is (in line with what Mark Adams wrote) that it is better to
> leave it to figure it out automatically. Also when using it as the coarse
> level preconditioner in a PCMG.
>
>
>
>
>
> --
>
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>



-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150917/c63c753e/attachment-0001.html>